A crisis communication network based on embodied conversational agents system with mobile services.
International Journal of Information Technology Volume 3 Number 4
A Crisis Communication Network Based on
Embodied Conversational Agents System with
Mobile Services
Ong Sing Goh, Cemal Ardil, Chun Che Fung, Kok Wai Wong, Arnold Depickere
ECA called Artificial Intelligence Neural-network Identity
(AINI) which delivers essential contents of news grabbed
from trusted first sources online documents and the
application of the ECA in mobile services.
The purpose of AINI is to deliver essential information from
trusted and updated sources and it is able to interact with its
users by ECA. The idea is to rely on a human-like
communication approach thereby providing a sense of comfort
and familiarity.
In this paper, one of the aims is to extend our basic AINI
framework to facilitate delivering content through mobile
services. Mobile communication technologies reduce reliance
on static communication methods (e.g. land-line phones), and
increase confidence and perceived safety when moving[1].
The portability of new miniaturized devices, together with
their ability to connect conveniently to networks in different
places, makes mobile computing possible. As mobile services
will be becoming more and more important, this has been the
motivation of extending the basic framework to include
mobile services.
Despite the recent growth in information and
communication technology, many applications have minimum
amount of intelligence in aid of human-computer
communication. On the other hand, practical Artificial
Intelligence (AI) technologies have gained wider acceptance
and have been incorporated in many information technology
(IT) applications. The aim of the development of the
conversation robot is to provide a human-like communication
environment. Such “humanized” communication approach is
the current trend in the IT world regardless whether it is webbased or mobile-based.
To achieve the above objective, an intelligent agent
software robot AINI has been developed. AINI has
customized Artificial Intelligence Markup Language
(AIML)[8] servable knowledge base being incorporated to
serve as a real conversation software robot in the CCNet
Portal. With the increasing availability of wireless and mobile
technologies, the proposed CCNet portal also uses the latest
technologies such as mobile chat and PDA chat, which in turn
will be used to send text-based information and images of the
latest information to subscribed users.
Abstract—In this paper, we proposed a new framework to
incorporate an intelligent agent software robot into a crisis
communication portal (CCNet) in order to send alert news to
subscribed users via email and other mobile services such as Short
Message Service (SMS), Multimedia Messaging Service (MMS) and
General Packet Radio Services (GPRS). The content on the mobile
services can be delivered either through mobile phone or Personal
Digital Assistance (PDA). This research has shown that with our
proposed framework, the embodied conversation agents system can
handle questions intelligently with our multilayer architecture. At the
same time, the extended framework can take care of delivery content
through a more humanoid interface on mobile devices.
Keywords—Crisis Communication Network (CCNet), Embodied
Conversational Agents (ECAs), Mobile Services, Artificial
Intelligence Neural-network Identity (AINI)
I. INTRODUCTION
D
URING the past decade, rapid advances in embodied
conversational agents, spoken language technology,
natural language processing, multimodal interfaces and
mobile applications have stimulated interest in a new class of
conversational interfaces [2], [3], [4], [5] and [6]. Many
researchers have been involved in AI researches into natural
language conversation [7], [8], [9], [10]. They have proposed
different techniques and produced several natural language
conversation systems. Every year they present their work by
competing for the Turing Test [11].
This paper aims to address the issues of managing global crisis
communication by introducing a crisis communication portal
called Crisis Communication Network (CCNet). In particular,
this paper focuses on two aspects of the system – an
Manuscript received June 2006. This research funded by Research Excellence
Grants Scheme, Murdoch University, Western Australia
Ong Sing Goh is with the Murdoch University, Perth, Western Australia
(e-mail: os.goh@murdoch.edu.au)
Cemal Ardil is with the National Academy of Azerbaijan, Baku,
Azerbaijan (e-mail: cemalardil@gmail.com)
Chun Che Fung is with the Murdoch University, Perth, Western Australia
(e-mail: l.fung@murdoch.edu.au)
Kok Wai Wong is with the Murdoch University, Perth, Western Australia
(e-mail: k.wong@murdoch.edu.au)
Arnold Depickere is with the Murdoch University, Perth, Western Australia
(e-mail: a.depickere@murdoch.edu.au).
257
International Journal of Information Technology Volume 3 Number 4
Client Layer
Data Layer
Application Layer
Reply
Query
Level 0
Spell Checker
Dictionary
Natural Language Understanding & Reasoning (NLUR)
network-based
advanced reasoning
Mobile
Browsers
WAPChat
Domain-specific KB
WEB
Broswers
SMSChat
Full Parsing
NLUR
ontology
semantic
network
gazetteer
FAQChat
FAQChat
Metadata
Index
Metadata Index Search
Question
understanding
Mobile
Gateway
natural language
understanding
document/sentence
understanding
natural language
understanding
Level 2
Level 3
PDAChat
Metadata Source
AlertNews
content categorizing
page wrapping
Open-domain
(AAA KB)
Conversation
Logs
WEBChat
Automated Knowledge
Extraction Agent (AKEA)
WWW
web crawling
Pattern Matching & Case Base Reasoning (PMCBR)
Multilevel Natural Language Query
WAP
Browsers
query network
Reasoning
GSM
Interface
Level 1
Level 4
Annotated ALICE AIML (AAA), trained Loebner prize
knowledge base
(XML Source)
Supervised Machine Learning by Domain Expert
Level 5
Fig. 1 AINI’s Architecture in the CCNet Portal
A human user can communicate with the developed system
using typed natural language conversation. The embodied
conversation agent system will reply text-prompts or Text-toSpeech Synthesis together with appropriate facial-expressions.
For the purpose of this research, the application area
chosen for designing the conversation agent is primarily in the
context of SARs epidemic crisis using scripting and
incorporation of artificial intelligence.
As illustrated in Fig. 1, AINI adopts a hybrid architecture
that combines multi-domain knowledge bases, multimodal
interface and multilevel natural language query. Given a
question, AINI first performs question analysis by extracting
pertinent information to be used in query formulation, such as
Noun Phrases and Verb Phrases. AINI employs an Internet
three-tier, thin client architecture that may be configured to
work with any web application. It comprises of a data layer,
application layer and client layer. This Internet specific
architecture offers a flexible solution to the unique
implementation requirements of the AINI system.
II. AINI’s ARCHITECTURE
This research project involves the establishment of a CCNet
portal1. The objective is to use the ECA, called Artificial
Intelligent Neural-network Identity [12, 13] as the basic
architecture. Our real-time prototype relies on distributed
agent architecture designed specifically for the web and
mobile technology. The software agent is based on a
conversation engine using a multi-domain knowledge model
and with multimodal human-computer communication
interface. It also offers multilevel natural language query
which communicates with one another via TCP/IP. In short,
AINI is a conversation agent designed by the authors that is
capable of having a meaningful conversation with the users.
From another perspective, AINI can be considered as a
software conversation robot, which uses a form of humancomputer communication system which is a combination of
natural language processing and multimodal communication.
1
The experiment portal can be access at http://ainibot.murdoch.edu.au
258
International Journal of Information Technology Volume 3 Number 4
Fig. 2 Top-Down Nl-Query Approach
natural language processing and speech synthesized via a 3D
animated character known as avatar. AINI’s engine
implements its sophisticated decision making network based
on the information it encounters in the knowledge bases.
These decision-making capabilities make use of the XML
specifications. The input and output of each module is an
XML-encoded data structure that keeps track of the current
computational state. These modules are conceptualized as
transformations over this XML data structure. The system
accepts questions and requests from users, and processes the
queries based on the information contained in AINI’s
knowledge base.
The application server layer handles the processing of
logic and information requests. Here, one or more application
servers are configured to compute the dialogue logic through
the multilevel natural language query algorithm[17]. In this
layer we simulated goal-driven or top-down natural language
query (NL-Query) approach just like human’s process their
language. The top-down approach seems to be a good model
for explaining how humans use their knowledge in
conversation. From the literature search, we concluded that in
the field of Natural Language Processing (NLP), it seems that
the top-down approach is far the best approach. As shown in
the Fig. 2, our top-down NL-query approach consists of 6
level of queries, namely Spell Checker (Level 0), Fulldiscourse NLUR (Level 1), FAQChat (Level 2), Metadata
Index Search (Level 3), PMCBR (Level 4) and Supervised
Machine Learning by Domain Expert (Level 5).
A.. Data Layer
The data server layer serves as storage for permanent data
required by the system, where the epidemic knowledge bases
are stored. These databases are Dictionary, Domain-Specific,
Open Domain and conversation logs. The dictionary is an
ispell which was first ran on TOPS-20 systems at MIT-AI
lab2. Domain-Specific database is extracted by the Automated
Knowledge Extraction Agent (AKEA) which consists of Full
Parsing Natural Language Understanding and Reasoning
(NLUR), FAQChat and Metadata Index. AKEA was designed
to establish the knowledge base for a global crisis
communication system called CCNet. CCNet was proposed
during the height of the SARs epidemic in 2003[14].
The Open-Domain database is taken from the existing
award winner Turing Test. This trained Knowledge Base is
also called Annotated ALICE Artificial Intelligence Markup
Language (AAA) [8, 15] where the conversation logs reside.
These web-enabled databases are accessible via the SQL
query standard for database connectivity using MySQL
database.
B. Application Layer
The AINI Server and Mobile Gateway are located in the
application layer. WAP and SMS gateway[16] serve as mobile
gateway is used widely across the globe both for serving
millions of short messages (SMS) and pushing WAP services.
They function as the interconnection path between the client
layer and data layer in the CCNet Portal. AINI’s engine is a
unique intelligent agent framework. All communication with
AINI is performed through a natural interface that uses a
2
http://www.mit.edu/afs/sipb/project/sipb-athena/src/ispell/
C. Client Layer
The user interface resides in the thin-client layer and is
supporting web-based and mobile service interface. For the
259
International Journal of Information Technology Volume 3 Number 4
performance and lightweight client runtime that delivers
powerful and consistent user experiences across major
operating systems, browsers, mobile phones and devices. For
the WAP services, the application was embedded WAP
browsers from vendors such as Openwave3 and Nokia4.
The conversation engine is implemented by open-source
architecture employing Kannel Mobile gateway [16], PHP,
Perl scripting language, Apache Server and knowledge base
stored in a MySQL server.
web-based, its employs Multimodal Agent Markup Language
(MAML) interpreter or Microsoft SAPI to handle the user
interface. MAML is a prototype multimodal markup language
based on XML that enables animated presentation agents or
avatars. It involves a talking virtual lifelike 3D agent character
that is capable of having a fairly meaningful conversation.
However for mobile devices as they have small screens,
there are limitations on the amount of information that they
can be presented at one time. Reading large amounts of
information from such devices can require large amounts of
Fig. 3 AKEA Framework
scrolling and concentration. To reduce distraction,
interactions, and potential information overload, a better way
of presenting information might be through multilevel or
hierarchical mechanisms[18]. Hence, chatting mode interface
will be the better solution for mobile service. In addition,
current wireless network service vendors have introduced a
wide bandwidth telephone network, known as 3G
communication [19], and it enhances the possibility of
adapting a smartphone as a client in traditional distributed
systems. On the PDA or SmartPhone, our system required
Mobile Flash Player [20]. This Flash player is high
III. AINI’s DOMAIN KNOWLEDGE MODEL
In our research, the domain model is the taxonomy of
knowledge related to the topic of the presentation, or XMLlike metadata model. This will reduce the workload of the
domain administrator or domain expert to predict every input
typed by the user. Instead, this allows the author to put more
3
4
260
http://developer.openwave.com
http://www.nokiausa.com/support/software
International Journal of Information Technology Volume 3 Number 4
gazetteer will be implemented as domain-dependent modular
components, allowing future improvements.
effort on scripting conversation within a specified domain or
conversation Domain-specific.
We believe that the ultimate conversational humancomputer interface uses and requires different kinds of
approaches. Therefore, we have been working to develop a
domain knowledge model for building conversation and
interactive systems. For example, according to S. Kshirsagar
and N. Magnenat-Thalmann [17], having a small conversation
about the weather requires a lot less resources than a
philosophical discussion about the meaning of life. In our
research, we defined our conversation system as a collective
of specific conversation units; every unit handles a specific
conversation between user and computer. In our case,
Domain-Specific knowledge base is created from the epidemic
online document from selected websites5.
Domain knowledge is one of the dimensions that
determines the focus or direction of a conversational system.
An Open-domain will practice techniques based on
probabilistic measures and has a wider range of information
source. For a system that focuses on certain domains, it is
more likely that the techniques are more logic-based and wellfounded, with relatively limited sources as compared to an
Open-domain. A domain-oriented conversational system deals
with questions under a Domain-specific environment, and can
be seen as a richer approach. This is because natural language
processing systems can exploit domain knowledge and
ontologies. Advanced reasoning such as providing
explanations for answers, generalizing questions, etc is
impossible in Open-domain systems. Open-domain question
answering systems need to deal with questions about nearly
everything and it is difficult to rely on ontological information
due to the absence of wide and yet detailed world knowledge.
On the other hand, these systems have much more data to
exploit in the process of extracting the answers.
In our architecture, we have implemented a multi-domain
model: an Open-domain knowledge base which is converted
from the AAA knowledge base and a Domain-specific
knowledge base. The Domain-specific knowledge base is the
epidemic online document extracted by the AKEA. AKEA
will be discussed in more details in following section.
However, if the user converses out of the presentation topic,
we define this domain category as the Unanswered-domain
which the knowledge is currently not available and randomly
generated. This is to determine whether the user is chatting
within the domain of the presentation topic or the user is
conversing differently from the domain knowledge model
presented. By doing this, we have rectified the trait of the
conversation agent or software robot, from a diverse
conversation to a specific presentation topic. The web
knowledge base is continuously updated with facts extracted
from online epidemic news using information extraction (IE)
and knowledge representation by AKEA. IE is the task of
extracting relevant fragments of text from larger documents, to
allow the fragments to be processed further in some automated
way. For example, to answer a user’s query, the ontology and
5
A. Automated Knowledge Extraction Agent (AKEA)
The Fig 3. shows the architecture of the CCNet knowledge
extraction agent framework called AKEA. Four modules make
up the agent with the crawler as the interface between the
agent and the web. The crawler is like those used in
conventional crawler-based search engines. The crawler
resolves root domain names and follows subsequent links that
is available on a page until a certain depth is defined by the
user. These configurations are set in the crawler config
database. For every page crawled, a copy is returned for
further processing by the wrapper. The activities of the
crawler are logged in by the crawler log database.
Online news documents returned by the crawler are in the
hypertext format and consist of a variety of unwanted
characters (Fig. 4). The wrapper prepares the raw news by
separating the actual news content and other meta-information
from hypertext characters. This process is also known as
cleaning and the result is referred to as cleaned news (Fig. 5).
Information such as date of news, news title, news content and
many more is extracted and stored in the CCNet news
repository.
New meningitis threat being contained by web
of partnerships
...
8 APRIL 2004 | GENEVA -- A rare strain of
meningitis, which re-emerged recently in Burkina
Faso…
…
Fig. 4 Example of online news returned by crawler
title: New meningitis threat being contained by web
of partnerships
url:
http://www.who.int/mediacentre/releases/2004/pr25/en
/
date: 8 April 2004
content: A rare strain of meningitis, which reemerged recently in Burkina Faso…
Fig. 5 Example of cleaned news returned by wrapper
The syntactic preprocessor performs the task of
identifying the dependencies among words. Based on the
dependencies, grammatical relationship (i.e. phrasal
categories) like noun phrases, verb phrases and prepositional
phrases are extracted using sentence parser for the English
language like Link Grammar[21] and Minipar [22]. The
named entities in noun phrases are assigned with tags such as
disease, location and person using the weighted gazetteer
approach. A reference list known as gazetteer is used by the
preprocessor. These tags enable the agent to identify what type
of entity the corresponding noun phrases are and in which
level and node do these entities belong to in the ontology.
http://www.info.gov.hk/info/sars/
http://www.sars.gov.sg
261
International Journal of Information Technology Volume 3 Number 4
Pronouns are also resolved whenever necessary. The named
entities that have been tagged are inserted into the
corresponding entry in the news repository. For example,
using the cleaned news in Fig. 5, the syntactic preprocessor
managed to identify two named entities namely meningitis and
Burkina Faso. Using the gazetteer, the preprocessor will
discover that meningitis is a type of disease and Burkina Faso
is a country and tag them respectively using the ontology tag
in the form of named_entity [ontology tag].
In the gazetteers, each entry has additional information
like
weight,
ontology
id
and
the
acceptable
preceding/foregoing grammatical relations in addition to the
triggering information, category and entity name. For
example, a returned noun phrase “Japanese Encephalitis
disease”, could trigger ambiguity. This could be resolved by
just using the weighting mechanism without the need for any
hand-crafted rules.
N-gen
N-
ε
2
N-obj
ε
ε
N-
ε
ε
B
N-
H
K
N-
J
G
http://w
ww.who.
int/media
centre/rel
eases/20
04/pr25/
en/
N-
D
N-
N-
I
N-nn
A
ε
N-obj2
C
N-obj2
E
ε
N-s
8
April
2004
New
meningitis
threat being
contained by
web of
partnerships
A rare strain of
meningitis,
which reemerged
recently in
Burkina Faso…
meningitis[dis
ease] Burkina
Faso[country]
fields in the template namely the wh-token corresponding to
the ontology tag, first two lines of content, disease named
entity and URL. The first and second requires some
processing prior to replacement.
The ontology tag associated with each named entity is
resolved to obtain the corresponding wh-token. Currently, the
agent is capable of handling four types of wh-token: where,
resolved from location named-entities, when, resolved from
date named-entities, who, resolved from agent named-entities
and what. The what token is resolvable from all ontological
entities with additional tokens. For example, given the named
entity Burkina Faso and its tag country, we can obtain the
where token and what token with the country tag. This is
possible because the question where does meningitis…? is
similar to asking what country does meningitis…? As shown
in the Fig. 8 and Fig. 9.
A-
ε
L
URL
TABLE I
A SAMPLE ENTRY IN THE NEWS REPOSITORY
NAME
DATE
TITLE
CONTENT
ENTITY
N-
F
Fig. 6 Finite-state Automaton for noun phrases extraction
The algorithm adopted by our named-entity tagger
employs finite-state automaton. The sentence to be namedentity tagged is first parsed for syntactic categories and
grammatical relationships using the sentence parser of choice,
Minipar. The output of parsing is then fed through the Finitestate Automaton (FSA) as shown in Fig. 6 to extract noun
phrases. By inferring named-entity extraction criteria by [23],
all named-entity are subsets of noun phrases and not merely
proper nouns.
The information in the news repository is fed into two main
components, namely the CCNet portal and the AIML
converter. Information in the news repository is directly
published to the CCNet portal without any further processing.
Table 1 shows a sample entry in the news repository.
where _ meningitis _
A rare strain of meningitis, which reemerged recently in Burkina Faso…
window.open(“http://www.who.int/mediacentre/
releases/2004/pr25/en/”,””,””);
A Crisis Communication Network Based on
Embodied Conversational Agents System with
Mobile Services
Ong Sing Goh, Cemal Ardil, Chun Che Fung, Kok Wai Wong, Arnold Depickere
ECA called Artificial Intelligence Neural-network Identity
(AINI) which delivers essential contents of news grabbed
from trusted first sources online documents and the
application of the ECA in mobile services.
The purpose of AINI is to deliver essential information from
trusted and updated sources and it is able to interact with its
users by ECA. The idea is to rely on a human-like
communication approach thereby providing a sense of comfort
and familiarity.
In this paper, one of the aims is to extend our basic AINI
framework to facilitate delivering content through mobile
services. Mobile communication technologies reduce reliance
on static communication methods (e.g. land-line phones), and
increase confidence and perceived safety when moving[1].
The portability of new miniaturized devices, together with
their ability to connect conveniently to networks in different
places, makes mobile computing possible. As mobile services
will be becoming more and more important, this has been the
motivation of extending the basic framework to include
mobile services.
Despite the recent growth in information and
communication technology, many applications have minimum
amount of intelligence in aid of human-computer
communication. On the other hand, practical Artificial
Intelligence (AI) technologies have gained wider acceptance
and have been incorporated in many information technology
(IT) applications. The aim of the development of the
conversation robot is to provide a human-like communication
environment. Such “humanized” communication approach is
the current trend in the IT world regardless whether it is webbased or mobile-based.
To achieve the above objective, an intelligent agent
software robot AINI has been developed. AINI has
customized Artificial Intelligence Markup Language
(AIML)[8] servable knowledge base being incorporated to
serve as a real conversation software robot in the CCNet
Portal. With the increasing availability of wireless and mobile
technologies, the proposed CCNet portal also uses the latest
technologies such as mobile chat and PDA chat, which in turn
will be used to send text-based information and images of the
latest information to subscribed users.
Abstract—In this paper, we proposed a new framework to
incorporate an intelligent agent software robot into a crisis
communication portal (CCNet) in order to send alert news to
subscribed users via email and other mobile services such as Short
Message Service (SMS), Multimedia Messaging Service (MMS) and
General Packet Radio Services (GPRS). The content on the mobile
services can be delivered either through mobile phone or Personal
Digital Assistance (PDA). This research has shown that with our
proposed framework, the embodied conversation agents system can
handle questions intelligently with our multilayer architecture. At the
same time, the extended framework can take care of delivery content
through a more humanoid interface on mobile devices.
Keywords—Crisis Communication Network (CCNet), Embodied
Conversational Agents (ECAs), Mobile Services, Artificial
Intelligence Neural-network Identity (AINI)
I. INTRODUCTION
D
URING the past decade, rapid advances in embodied
conversational agents, spoken language technology,
natural language processing, multimodal interfaces and
mobile applications have stimulated interest in a new class of
conversational interfaces [2], [3], [4], [5] and [6]. Many
researchers have been involved in AI researches into natural
language conversation [7], [8], [9], [10]. They have proposed
different techniques and produced several natural language
conversation systems. Every year they present their work by
competing for the Turing Test [11].
This paper aims to address the issues of managing global crisis
communication by introducing a crisis communication portal
called Crisis Communication Network (CCNet). In particular,
this paper focuses on two aspects of the system – an
Manuscript received June 2006. This research funded by Research Excellence
Grants Scheme, Murdoch University, Western Australia
Ong Sing Goh is with the Murdoch University, Perth, Western Australia
(e-mail: os.goh@murdoch.edu.au)
Cemal Ardil is with the National Academy of Azerbaijan, Baku,
Azerbaijan (e-mail: cemalardil@gmail.com)
Chun Che Fung is with the Murdoch University, Perth, Western Australia
(e-mail: l.fung@murdoch.edu.au)
Kok Wai Wong is with the Murdoch University, Perth, Western Australia
(e-mail: k.wong@murdoch.edu.au)
Arnold Depickere is with the Murdoch University, Perth, Western Australia
(e-mail: a.depickere@murdoch.edu.au).
257
International Journal of Information Technology Volume 3 Number 4
Client Layer
Data Layer
Application Layer
Reply
Query
Level 0
Spell Checker
Dictionary
Natural Language Understanding & Reasoning (NLUR)
network-based
advanced reasoning
Mobile
Browsers
WAPChat
Domain-specific KB
WEB
Broswers
SMSChat
Full Parsing
NLUR
ontology
semantic
network
gazetteer
FAQChat
FAQChat
Metadata
Index
Metadata Index Search
Question
understanding
Mobile
Gateway
natural language
understanding
document/sentence
understanding
natural language
understanding
Level 2
Level 3
PDAChat
Metadata Source
AlertNews
content categorizing
page wrapping
Open-domain
(AAA KB)
Conversation
Logs
WEBChat
Automated Knowledge
Extraction Agent (AKEA)
WWW
web crawling
Pattern Matching & Case Base Reasoning (PMCBR)
Multilevel Natural Language Query
WAP
Browsers
query network
Reasoning
GSM
Interface
Level 1
Level 4
Annotated ALICE AIML (AAA), trained Loebner prize
knowledge base
(XML Source)
Supervised Machine Learning by Domain Expert
Level 5
Fig. 1 AINI’s Architecture in the CCNet Portal
A human user can communicate with the developed system
using typed natural language conversation. The embodied
conversation agent system will reply text-prompts or Text-toSpeech Synthesis together with appropriate facial-expressions.
For the purpose of this research, the application area
chosen for designing the conversation agent is primarily in the
context of SARs epidemic crisis using scripting and
incorporation of artificial intelligence.
As illustrated in Fig. 1, AINI adopts a hybrid architecture
that combines multi-domain knowledge bases, multimodal
interface and multilevel natural language query. Given a
question, AINI first performs question analysis by extracting
pertinent information to be used in query formulation, such as
Noun Phrases and Verb Phrases. AINI employs an Internet
three-tier, thin client architecture that may be configured to
work with any web application. It comprises of a data layer,
application layer and client layer. This Internet specific
architecture offers a flexible solution to the unique
implementation requirements of the AINI system.
II. AINI’s ARCHITECTURE
This research project involves the establishment of a CCNet
portal1. The objective is to use the ECA, called Artificial
Intelligent Neural-network Identity [12, 13] as the basic
architecture. Our real-time prototype relies on distributed
agent architecture designed specifically for the web and
mobile technology. The software agent is based on a
conversation engine using a multi-domain knowledge model
and with multimodal human-computer communication
interface. It also offers multilevel natural language query
which communicates with one another via TCP/IP. In short,
AINI is a conversation agent designed by the authors that is
capable of having a meaningful conversation with the users.
From another perspective, AINI can be considered as a
software conversation robot, which uses a form of humancomputer communication system which is a combination of
natural language processing and multimodal communication.
1
The experiment portal can be access at http://ainibot.murdoch.edu.au
258
International Journal of Information Technology Volume 3 Number 4
Fig. 2 Top-Down Nl-Query Approach
natural language processing and speech synthesized via a 3D
animated character known as avatar. AINI’s engine
implements its sophisticated decision making network based
on the information it encounters in the knowledge bases.
These decision-making capabilities make use of the XML
specifications. The input and output of each module is an
XML-encoded data structure that keeps track of the current
computational state. These modules are conceptualized as
transformations over this XML data structure. The system
accepts questions and requests from users, and processes the
queries based on the information contained in AINI’s
knowledge base.
The application server layer handles the processing of
logic and information requests. Here, one or more application
servers are configured to compute the dialogue logic through
the multilevel natural language query algorithm[17]. In this
layer we simulated goal-driven or top-down natural language
query (NL-Query) approach just like human’s process their
language. The top-down approach seems to be a good model
for explaining how humans use their knowledge in
conversation. From the literature search, we concluded that in
the field of Natural Language Processing (NLP), it seems that
the top-down approach is far the best approach. As shown in
the Fig. 2, our top-down NL-query approach consists of 6
level of queries, namely Spell Checker (Level 0), Fulldiscourse NLUR (Level 1), FAQChat (Level 2), Metadata
Index Search (Level 3), PMCBR (Level 4) and Supervised
Machine Learning by Domain Expert (Level 5).
A.. Data Layer
The data server layer serves as storage for permanent data
required by the system, where the epidemic knowledge bases
are stored. These databases are Dictionary, Domain-Specific,
Open Domain and conversation logs. The dictionary is an
ispell which was first ran on TOPS-20 systems at MIT-AI
lab2. Domain-Specific database is extracted by the Automated
Knowledge Extraction Agent (AKEA) which consists of Full
Parsing Natural Language Understanding and Reasoning
(NLUR), FAQChat and Metadata Index. AKEA was designed
to establish the knowledge base for a global crisis
communication system called CCNet. CCNet was proposed
during the height of the SARs epidemic in 2003[14].
The Open-Domain database is taken from the existing
award winner Turing Test. This trained Knowledge Base is
also called Annotated ALICE Artificial Intelligence Markup
Language (AAA) [8, 15] where the conversation logs reside.
These web-enabled databases are accessible via the SQL
query standard for database connectivity using MySQL
database.
B. Application Layer
The AINI Server and Mobile Gateway are located in the
application layer. WAP and SMS gateway[16] serve as mobile
gateway is used widely across the globe both for serving
millions of short messages (SMS) and pushing WAP services.
They function as the interconnection path between the client
layer and data layer in the CCNet Portal. AINI’s engine is a
unique intelligent agent framework. All communication with
AINI is performed through a natural interface that uses a
2
http://www.mit.edu/afs/sipb/project/sipb-athena/src/ispell/
C. Client Layer
The user interface resides in the thin-client layer and is
supporting web-based and mobile service interface. For the
259
International Journal of Information Technology Volume 3 Number 4
performance and lightweight client runtime that delivers
powerful and consistent user experiences across major
operating systems, browsers, mobile phones and devices. For
the WAP services, the application was embedded WAP
browsers from vendors such as Openwave3 and Nokia4.
The conversation engine is implemented by open-source
architecture employing Kannel Mobile gateway [16], PHP,
Perl scripting language, Apache Server and knowledge base
stored in a MySQL server.
web-based, its employs Multimodal Agent Markup Language
(MAML) interpreter or Microsoft SAPI to handle the user
interface. MAML is a prototype multimodal markup language
based on XML that enables animated presentation agents or
avatars. It involves a talking virtual lifelike 3D agent character
that is capable of having a fairly meaningful conversation.
However for mobile devices as they have small screens,
there are limitations on the amount of information that they
can be presented at one time. Reading large amounts of
information from such devices can require large amounts of
Fig. 3 AKEA Framework
scrolling and concentration. To reduce distraction,
interactions, and potential information overload, a better way
of presenting information might be through multilevel or
hierarchical mechanisms[18]. Hence, chatting mode interface
will be the better solution for mobile service. In addition,
current wireless network service vendors have introduced a
wide bandwidth telephone network, known as 3G
communication [19], and it enhances the possibility of
adapting a smartphone as a client in traditional distributed
systems. On the PDA or SmartPhone, our system required
Mobile Flash Player [20]. This Flash player is high
III. AINI’s DOMAIN KNOWLEDGE MODEL
In our research, the domain model is the taxonomy of
knowledge related to the topic of the presentation, or XMLlike metadata model. This will reduce the workload of the
domain administrator or domain expert to predict every input
typed by the user. Instead, this allows the author to put more
3
4
260
http://developer.openwave.com
http://www.nokiausa.com/support/software
International Journal of Information Technology Volume 3 Number 4
gazetteer will be implemented as domain-dependent modular
components, allowing future improvements.
effort on scripting conversation within a specified domain or
conversation Domain-specific.
We believe that the ultimate conversational humancomputer interface uses and requires different kinds of
approaches. Therefore, we have been working to develop a
domain knowledge model for building conversation and
interactive systems. For example, according to S. Kshirsagar
and N. Magnenat-Thalmann [17], having a small conversation
about the weather requires a lot less resources than a
philosophical discussion about the meaning of life. In our
research, we defined our conversation system as a collective
of specific conversation units; every unit handles a specific
conversation between user and computer. In our case,
Domain-Specific knowledge base is created from the epidemic
online document from selected websites5.
Domain knowledge is one of the dimensions that
determines the focus or direction of a conversational system.
An Open-domain will practice techniques based on
probabilistic measures and has a wider range of information
source. For a system that focuses on certain domains, it is
more likely that the techniques are more logic-based and wellfounded, with relatively limited sources as compared to an
Open-domain. A domain-oriented conversational system deals
with questions under a Domain-specific environment, and can
be seen as a richer approach. This is because natural language
processing systems can exploit domain knowledge and
ontologies. Advanced reasoning such as providing
explanations for answers, generalizing questions, etc is
impossible in Open-domain systems. Open-domain question
answering systems need to deal with questions about nearly
everything and it is difficult to rely on ontological information
due to the absence of wide and yet detailed world knowledge.
On the other hand, these systems have much more data to
exploit in the process of extracting the answers.
In our architecture, we have implemented a multi-domain
model: an Open-domain knowledge base which is converted
from the AAA knowledge base and a Domain-specific
knowledge base. The Domain-specific knowledge base is the
epidemic online document extracted by the AKEA. AKEA
will be discussed in more details in following section.
However, if the user converses out of the presentation topic,
we define this domain category as the Unanswered-domain
which the knowledge is currently not available and randomly
generated. This is to determine whether the user is chatting
within the domain of the presentation topic or the user is
conversing differently from the domain knowledge model
presented. By doing this, we have rectified the trait of the
conversation agent or software robot, from a diverse
conversation to a specific presentation topic. The web
knowledge base is continuously updated with facts extracted
from online epidemic news using information extraction (IE)
and knowledge representation by AKEA. IE is the task of
extracting relevant fragments of text from larger documents, to
allow the fragments to be processed further in some automated
way. For example, to answer a user’s query, the ontology and
5
A. Automated Knowledge Extraction Agent (AKEA)
The Fig 3. shows the architecture of the CCNet knowledge
extraction agent framework called AKEA. Four modules make
up the agent with the crawler as the interface between the
agent and the web. The crawler is like those used in
conventional crawler-based search engines. The crawler
resolves root domain names and follows subsequent links that
is available on a page until a certain depth is defined by the
user. These configurations are set in the crawler config
database. For every page crawled, a copy is returned for
further processing by the wrapper. The activities of the
crawler are logged in by the crawler log database.
Online news documents returned by the crawler are in the
hypertext format and consist of a variety of unwanted
characters (Fig. 4). The wrapper prepares the raw news by
separating the actual news content and other meta-information
from hypertext characters. This process is also known as
cleaning and the result is referred to as cleaned news (Fig. 5).
Information such as date of news, news title, news content and
many more is extracted and stored in the CCNet news
repository.
New meningitis threat being contained by web
of partnerships
...
8 APRIL 2004 | GENEVA -- A rare strain of
meningitis, which re-emerged recently in Burkina
Faso…
…
Fig. 4 Example of online news returned by crawler
title: New meningitis threat being contained by web
of partnerships
url:
http://www.who.int/mediacentre/releases/2004/pr25/en
/
date: 8 April 2004
content: A rare strain of meningitis, which reemerged recently in Burkina Faso…
Fig. 5 Example of cleaned news returned by wrapper
The syntactic preprocessor performs the task of
identifying the dependencies among words. Based on the
dependencies, grammatical relationship (i.e. phrasal
categories) like noun phrases, verb phrases and prepositional
phrases are extracted using sentence parser for the English
language like Link Grammar[21] and Minipar [22]. The
named entities in noun phrases are assigned with tags such as
disease, location and person using the weighted gazetteer
approach. A reference list known as gazetteer is used by the
preprocessor. These tags enable the agent to identify what type
of entity the corresponding noun phrases are and in which
level and node do these entities belong to in the ontology.
http://www.info.gov.hk/info/sars/
http://www.sars.gov.sg
261
International Journal of Information Technology Volume 3 Number 4
Pronouns are also resolved whenever necessary. The named
entities that have been tagged are inserted into the
corresponding entry in the news repository. For example,
using the cleaned news in Fig. 5, the syntactic preprocessor
managed to identify two named entities namely meningitis and
Burkina Faso. Using the gazetteer, the preprocessor will
discover that meningitis is a type of disease and Burkina Faso
is a country and tag them respectively using the ontology tag
in the form of named_entity [ontology tag].
In the gazetteers, each entry has additional information
like
weight,
ontology
id
and
the
acceptable
preceding/foregoing grammatical relations in addition to the
triggering information, category and entity name. For
example, a returned noun phrase “Japanese Encephalitis
disease”, could trigger ambiguity. This could be resolved by
just using the weighting mechanism without the need for any
hand-crafted rules.
N-gen
N-
ε
2
N-obj
ε
ε
N-
ε
ε
B
N-
H
K
N-
J
G
http://w
ww.who.
int/media
centre/rel
eases/20
04/pr25/
en/
N-
D
N-
N-
I
N-nn
A
ε
N-obj2
C
N-obj2
E
ε
N-s
8
April
2004
New
meningitis
threat being
contained by
web of
partnerships
A rare strain of
meningitis,
which reemerged
recently in
Burkina Faso…
meningitis[dis
ease] Burkina
Faso[country]
fields in the template namely the wh-token corresponding to
the ontology tag, first two lines of content, disease named
entity and URL. The first and second requires some
processing prior to replacement.
The ontology tag associated with each named entity is
resolved to obtain the corresponding wh-token. Currently, the
agent is capable of handling four types of wh-token: where,
resolved from location named-entities, when, resolved from
date named-entities, who, resolved from agent named-entities
and what. The what token is resolvable from all ontological
entities with additional tokens. For example, given the named
entity Burkina Faso and its tag country, we can obtain the
where token and what token with the country tag. This is
possible because the question where does meningitis…? is
similar to asking what country does meningitis…? As shown
in the Fig. 8 and Fig. 9.
A-
ε
L
URL
TABLE I
A SAMPLE ENTRY IN THE NEWS REPOSITORY
NAME
DATE
TITLE
CONTENT
ENTITY
N-
F
Fig. 6 Finite-state Automaton for noun phrases extraction
The algorithm adopted by our named-entity tagger
employs finite-state automaton. The sentence to be namedentity tagged is first parsed for syntactic categories and
grammatical relationships using the sentence parser of choice,
Minipar. The output of parsing is then fed through the Finitestate Automaton (FSA) as shown in Fig. 6 to extract noun
phrases. By inferring named-entity extraction criteria by [23],
all named-entity are subsets of noun phrases and not merely
proper nouns.
The information in the news repository is fed into two main
components, namely the CCNet portal and the AIML
converter. Information in the news repository is directly
published to the CCNet portal without any further processing.
Table 1 shows a sample entry in the news repository.
where _ meningitis _
A rare strain of meningitis, which reemerged recently in Burkina Faso…
window.open(“http://www.who.int/mediacentre/
releases/2004/pr25/en/”,””,””);